Gamma Log Inversion of Seismic Data Based on Transformer With Stratigraphic Position Encoding

As an indispensable part of geophysical exploration, seismic inversion can obtain the properties of subsurface media based on seismic data and available well-log information. With the nonlinear mapping ability, deep neural networks can map seismic data to well-log of interest. Interpreting gamma is crucial as it is essential for determining lithology and indicating sediment characteristics. Stratigraphic frameworks can approximate low-frequency trends in subsurface properties and are often used to guide well-log interpolation effectively. However, the existing deep neural network models cannot effectively explicitly fuse critical stratigraphic information, which will restrict the physical explainability and correctness of the seismic inversion. Thus, we propose a stratigraphic-encoded transformer algorithm, named SeisWellTrans, to build a gamma log inversion model using horizon position encoding and seismic trace as inputs. Specifically, the incorporation of stratigraphic information from several horizons is crucial for improving the resolution of the output; and SeisWellTrans can efficiently model context in seismic sequences by capturing the interactions between horizon position encodings. We take the Volve field data as an example and use several gamma curves as training labels, and numerical experiments demonstrate the geologically reasonable performance and high validation accuracy of this network and the crucial role that stratigraphic information plays. On the four validation wells, stratigraphic-encoded SeisWellTrans obtained an average correlation coefficient of 86%, exceeding 79% of stratigraphic-encoded convolutional neural network (CNN).